Future deprecation notice
=========================
A near-future release of OpenSSH will switch scp(1) from using the
legacy scp/rcp protocol to using SFTP by default.
Legacy scp/rcp performs wildcard expansion of remote filenames (e.g.
"scp host:* .") through the remote shell. This has the side effect of
requiring double quoting of shell meta-characters in file names
included on scp(1) command-lines, otherwise they could be interpreted
as shell commands on the remote side.
This creates one area of potential incompatibility: scp(1) when using
the SFTP protocol no longer requires this finicky and brittle quoting,
and attempts to use it may cause transfers to fail. We consider the
removal of the need for double-quoting shell characters in file names
to be a benefit and do not intend to introduce bug-compatibility for
legacy scp/rcp in scp(1) when using the SFTP protocol.
Another area of potential incompatibility relates to the use of remote
paths relative to other user's home directories, for example -
"scp host:~user/file /tmp". The SFTP protocol has no native way to
expand a ~user path. However, sftp-server(8) in OpenSSH 8.7 and later
support a protocol extension "expand-path@openssh.com" to support
this.
Security Near Miss
==================
* sshd(8): fix an integer overflow in the user authentication path
that, in conjunction with other logic errors, could have yielded
unauthenticated access under difficult to exploit conditions.
This situation is not exploitable because of independent checks in
the privilege separation monitor. Privilege separation has been
enabled by default in since openssh-3.2.2 (released in 2002) and
has been mandatory since openssh-7.5 (released in 2017). Moreover,
portable OpenSSH has used toolchain features available in most
modern compilers to abort on signed integer overflow since
openssh-6.5 (released in 2014).
Thanks to Malcolm Stagg for finding and reporting this bug.
Potentially-incompatible changes
================================
* sshd(8), portable OpenSSH only: this release removes in-built
support for MD5-hashed passwords. If you require these on your
system then we recommend linking against libxcrypt or similar.
* This release modifies the FIDO security key middleware interface
and increments SSH_SK_VERSION_MAJOR.
Changes since OpenSSH 8.8
=========================
This release includes a number of new features.
New features
------------
* ssh(1), sshd(8), ssh-add(1), ssh-agent(1): add a system for
restricting forwarding and use of keys added to ssh-agent(1)
A detailed description of the feature is available at
https://www.openssh.com/agent-restrict.html and the protocol
extensions are documented in the PROTOCOL and PROTOCOL.agent
files in the source release.
* ssh(1), sshd(8): add the sntrup761x25519-sha512@openssh.com hybrid
ECDH/x25519 + Streamlined NTRU Prime post-quantum KEX to the
default KEXAlgorithms list (after the ECDH methods but before the
prime-group DH ones). The next release of OpenSSH is likely to
make this key exchange the default method.
* ssh-keygen(1): when downloading resident keys from a FIDO token,
pass back the user ID that was used when the key was created and
append it to the filename the key is written to (if it is not the
default). Avoids keys being clobbered if the user created multiple
resident keys with the same application string but different user
IDs.
* ssh-keygen(1), ssh(1), ssh-agent(1): better handling for FIDO keys
on tokens that provide user verification (UV) on the device itself,
including biometric keys, avoiding unnecessary PIN prompts.
* ssh-keygen(1): add "ssh-keygen -Y match-principals" operation to
perform matching of principals names against an allowed signers
file. To be used towards a TOFU model for SSH signatures in git.
* ssh-add(1), ssh-agent(1): allow pin-required FIDO keys to be added
to ssh-agent(1). $SSH_ASKPASS will be used to request the PIN at
authentication time.
* ssh-keygen(1): allow selection of hash at sshsig signing time
(either sha512 (default) or sha256).
* ssh(1), sshd(8): read network data directly to the packet input
buffer instead indirectly via a small stack buffer. Provides a
modest performance improvement.
* ssh(1), sshd(8): read data directly to the channel input buffer,
providing a similar modest performance improvement.
* ssh(1): extend the PubkeyAuthentication configuration directive to
accept yes|no|unbound|host-bound to allow control over one of the
protocol extensions used to implement agent-restricted keys.
Bugfixes
--------
* sshd(8): document that CASignatureAlgorithms, ExposeAuthInfo and
PubkeyAuthOptions can be used in a Match block. PR#277.
* sshd(8): fix possible string truncation when constructing paths to
.rhosts/.shosts files with very long user home directory names.
* ssh-keysign(1): unbreak for KEX algorithms that use SHA384/512
exchange hashes
* ssh(1): don't put the TTY into raw mode when SessionType=none,
avoids ^C being unable to kill such a session. bz3360
* scp(1): fix some corner-case bugs in SFTP-mode handling of
~-prefixed paths.
* ssh(1): unbreak hostbased auth using RSA keys. Allow ssh(1) to
select RSA keys when only RSA/SHA2 signature algorithms are
configured (this is the default case). Previously RSA keys were
not being considered in the default case.
* ssh-keysign(1): make ssh-keysign use the requested signature
algorithm and not the default for the key type. Part of unbreaking
hostbased auth for RSA/SHA2 keys.
* ssh(1): stricter UpdateHostkey signature verification logic on
the client- side. Require RSA/SHA2 signatures for RSA hostkeys
except when RSA/SHA1 was explicitly negotiated during initial
KEX; bz3375
* ssh(1), sshd(8): fix signature algorithm selection logic for
UpdateHostkeys on the server side. The previous code tried to
prefer RSA/SHA2 for hostkey proofs of RSA keys, but missed some
cases. This will use RSA/SHA2 signatures for RSA keys if the
client proposed these algorithms in initial KEX. bz3375
* All: convert all uses of select(2)/pselect(2) to poll(2)/ppoll(2).
This includes the mainloops in ssh(1), ssh-agent(1), ssh-agent(1)
and sftp-server(8), as well as the sshd(8) listen loop and all
other FD read/writability checks. On platforms with missing or
broken poll(2)/ppoll(2) syscalls a select(2)-based compat shim is
available.
* ssh-keygen(1): the "-Y find-principals" command was verifying key
validity when using ca certs but not with simple key lifetimes
within the allowed signers file.
* ssh-keygen(1): make sshsig verify-time argument parsing optional
* sshd(8): fix truncation in rhosts/shosts path construction.
* ssh(1), ssh-agent(1): avoid xmalloc(0) for PKCS#11 keyid for ECDSA
keys (we already did this for RSA keys). Avoids fatal errors for
PKCS#11 libraries that return empty keyid, e.g. Microchip ATECC608B
"cryptoauthlib"; bz#3364
* ssh(1), ssh-agent(1): improve the testing of credentials against
inserted FIDO: ask the token whether a particular key belongs to
it in cases where the token supports on-token user-verification
(e.g. biometrics) rather than just assuming that it will accept it.
Will reduce spurious "Confirm user presence" notifications for key
handles that relate to FIDO keys that are not currently inserted in at
least some cases. bz3366
* ssh(1), sshd(8): correct value for IPTOS_DSCP_LE. It needs to
allow for the preceding two ECN bits. bz#3373
* ssh-keygen(1): add missing -O option to usage() for the "-Y sign"
option.
* ssh-keygen(1): fix a NULL deref when using the find-principals
function, when matching an allowed_signers line that contains a
namespace restriction, but no restriction specified on the
command-line
* ssh-agent(1): fix memleak in process_extension(); oss-fuzz
issue #42719
* ssh(1): suppress "Connection to xxx closed" messages when LogLevel
is set to "error" or above. bz3378
* ssh(1), sshd(8): use correct zlib flags when inflate(3)-ing
compressed packet data. bz3372
* scp(1): when recursively transferring files in SFTP mode, create the
destination directory if it doesn't already exist to match scp(1) in
legacy RCP mode behaviour.
* scp(1): many improvements in error message consistency between scp(1)
in SFTP mode vs legacy RCP mode.
* sshd(8): fix potential race in SIGTERM handling PR#289
* ssh(1), ssh(8): since DSA keys are deprecated, move them to the
end of the default list of public keys so that they will be tried
last. PR#295
* ssh-keygen(1): allow 'ssh-keygen -Y find-principals' to match
wildcard principals in allowed_signers files
Portability
-----------
* ssh(1), sshd(8): don't trust closefrom(2) on Linux. glibc's
implementation does not work in a chroot when the kernel does not
have close_range(2). It tries to read from /proc/self/fd and when
that fails dies with an assertion of sorts. Instead, call
close_range(2) directly from our compat code and fall back if
that fails. bz#3349,
* OS X poll(2) is broken; use compat replacement. For character-
special devices like /dev/null, Darwin's poll(2) returns POLLNVAL
when polled with POLLIN. Apparently this is Apple bug 3710161 -
not public but a websearch will find other OSS projects
rediscovering it periodically since it was first identified in
2005.
* Correct handling of exceptfds/POLLPRI in our select(2)-based
poll(2)/ppoll(2) compat implementation.
* Cygwin: correct checking of mbstowcs() return value.
* Add a basic SECURITY.md that refers people to the openssh.com
website.
* Enable additional compiler warnings and toolchain hardening flags,
including -Wbitwise-instead-of-logical, -Wmisleading-indentation,
-fzero-call-used-regs and -ftrivial-auto-var-init.
* HP/UX. Use compat getline(3) on HP-UX 10.x, where the libc version
is not reliable.
*) Avoid loading of a dynamic engine twice.
[Bernd Edlinger]
*) Fixed building on Debian with kfreebsd kernels
[Mattias Ellert]
*) Prioritise DANE TLSA issuer certs over peer certs
[Viktor Dukhovni]
*) Fixed random API for MacOS prior to 10.12
These MacOS versions don't support the CommonCrypto APIs
[Lenny Primak]
Changes between 1.1.1k and 1.1.1l [24 Aug 2021]
*) Fixed an SM2 Decryption Buffer Overflow.
In order to decrypt SM2 encrypted data an application is expected
to call the API function EVP_PKEY_decrypt(). Typically an application
will call this function twice. The first time, on entry, the "out"
parameter can be NULL and, on exit, the "outlen" parameter is
populated with the buffer size required to hold the decrypted
plaintext. The application can then allocate a sufficiently sized
buffer and call EVP_PKEY_decrypt() again, but this time passing
a non-NULL value for the "out" parameter.
A bug in the implementation of the SM2 decryption code means that
the calculation of the buffer size required to hold the plaintext
returned by the first call to EVP_PKEY_decrypt() can be smaller
than the actual size required by the second call. This can lead to
a buffer overflow when EVP_PKEY_decrypt() is called by the application
a second time with a buffer that is too small.
A malicious attacker who is able present SM2 content for decryption
to an application could cause attacker chosen data to overflow the
buffer by up to a maximum of 62 bytes altering the contents of
other data held after the buffer, possibly changing application
behaviour or causing the application to crash. The location of the
buffer is application dependent but is typically heap allocated.
(CVE-2021-3711)
[Matt Caswell]
*) Fixed various read buffer overruns processing ASN.1 strings
ASN.1 strings are represented internally within OpenSSL as an
ASN1_STRING structure which contains a buffer holding the string
data and a field holding the buffer length. This contrasts with
normal C strings which are repesented as a buffer for the string
data which is terminated with a NUL (0) byte.
Although not a strict requirement, ASN.1 strings that are parsed
using OpenSSL's own "d2i" functions (and other similar parsing
functions) as well as any string whose value has been set with the
ASN1_STRING_set() function will additionally NUL terminate the byte
array in the ASN1_STRING structure.
However, it is possible for applications to directly construct
valid ASN1_STRING structures which do not NUL terminate the byte
array by directly setting the "data" and "length" fields in the
ASN1_STRING array. This can also happen by using the ASN1_STRING_set0()
function.
Numerous OpenSSL functions that print ASN.1 data have been found
to assume that the ASN1_STRING byte array will be NUL terminated,
even though this is not guaranteed for strings that have been
directly constructed. Where an application requests an ASN.1
structure to be printed, and where that ASN.1 structure contains
ASN1_STRINGs that have been directly constructed by the application
without NUL terminating the "data" field, then a read buffer overrun
can occur.
The same thing can also occur during name constraints processing
of certificates (for example if a certificate has been directly
constructed by the application instead of loading it via the OpenSSL
parsing functions, and the certificate contains non NUL terminated
ASN1_STRING structures). It can also occur in the X509_get1_email(),
X509_REQ_get1_email() and X509_get1_ocsp() functions.
If a malicious actor can cause an application to directly construct
an ASN1_STRING and then process it through one of the affected
OpenSSL functions then this issue could be hit. This might result
in a crash (causing a Denial of Service attack). It could also
result in the disclosure of private memory contents (such as private
keys, or sensitive plaintext).
(CVE-2021-3712)
[Matt Caswell]
OpenSSH 8.7 has deprecated ChallengeResponseAuthentication, but not removed
it. It is now an alias for KbdInteractiveAuthentication (as are the prior
aliases of ChallengeResponseAuthentication).
I think this chunk was accidentally dropped in the OpenSSH 8.7 merge.
Imminent deprecation notice
===========================
OpenSSH will disable the ssh-rsa signature scheme by default in the
next release.
In the SSH protocol, the "ssh-rsa" signature scheme uses the SHA-1
hash algorithm in conjunction with the RSA public key algorithm.
It is now possible[1] to perform chosen-prefix attacks against the
SHA-1 algorithm for less than USD$50K.
Note that the deactivation of "ssh-rsa" signatures does not necessarily
require cessation of use for RSA keys. In the SSH protocol, keys may be
capable of signing using multiple algorithms. In particular, "ssh-rsa"
keys are capable of signing using "rsa-sha2-256" (RSA/SHA256),
"rsa-sha2-512" (RSA/SHA512) and "ssh-rsa" (RSA/SHA1). Only the last of
these is being turned off by default.
This algorithm is unfortunately still used widely despite the
existence of better alternatives, being the only remaining public key
signature algorithm specified by the original SSH RFCs that is still
enabled by default.
The better alternatives include:
* The RFC8332 RSA SHA-2 signature algorithms rsa-sha2-256/512. These
algorithms have the advantage of using the same key type as
"ssh-rsa" but use the safe SHA-2 hash algorithms. These have been
supported since OpenSSH 7.2 and are already used by default if the
client and server support them.
* The RFC8709 ssh-ed25519 signature algorithm. It has been supported
in OpenSSH since release 6.5.
* The RFC5656 ECDSA algorithms: ecdsa-sha2-nistp256/384/521. These
have been supported by OpenSSH since release 5.7.
To check whether a server is using the weak ssh-rsa public key
algorithm, for host authentication, try to connect to it after
removing the ssh-rsa algorithm from ssh(1)'s allowed list:
ssh -oHostKeyAlgorithms=-ssh-rsa user@host
If the host key verification fails and no other supported host key
types are available, the server software on that host should be
upgraded.
OpenSSH recently enabled the UpdateHostKeys option by default to
assist the client by automatically migrating to better algorithms.
[1] "SHA-1 is a Shambles: First Chosen-Prefix Collision on SHA-1 and
Application to the PGP Web of Trust" Leurent, G and Peyrin, T
(2020) https://eprint.iacr.org/2020/014.pdf
Potentially-incompatible changes
================================
This release includes a number of changes that may affect existing
configurations:
* scp(1): this release changes the behaviour of remote to remote
copies (e.g. "scp host-a:/path host-b:") to transfer through the
local host by default. This was previously available via the -3
flag. This mode avoids the need to expose credentials on the
origin hop, avoids triplicate interpretation of filenames by the
shell (by the local system, the copy origin and the destination)
and, in conjunction with the SFTP support for scp(1) mentioned
below, allows use of all authentication methods to the remote
hosts (previously, only non-interactive methods could be used).
A -R flag has been added to select the old behaviour.
* ssh(1)/sshd(8): both the client and server are now using a
stricter configuration file parser. The new parser uses more
shell-like rules for quotes, space and escape characters. It is
also more strict in rejecting configurations that include options
lacking arguments. Previously some options (e.g. DenyUsers) could
appear on a line with no subsequent arguments. This release will
reject such configurations. The new parser will also reject
configurations with unterminated quotes and multiple '='
characters after the option name.
* ssh(1): when using SSHFP DNS records for host key verification,
ssh(1) will verify all matching records instead of just those
with the specific signature type requested. This may cause host
key verification problems if stale SSHFP records of a different
or legacy signature type exist alongside other records for a
particular host. bz#3322
* ssh-keygen(1): when generating a FIDO key and specifying an
explicit attestation challenge (using -Ochallenge), the challenge
will now be hashed by the builtin security key middleware. This
removes the (undocumented) requirement that challenges be exactly
32 bytes in length and matches the expectations of libfido2.
* sshd(8): environment="..." directives in authorized_keys files are
now first-match-wins and limited to 1024 discrete environment
variable names.
Changes since OpenSSH 8.6
=========================
This release contains a mix of new features and bug-fixes.
New features
------------
- scp(1): experimental support for transfers using the SFTP protocol
as a replacement for the venerable SCP/RCP protocol that it has
traditionally used. SFTP offers more predictable filename handling
and does not require expansion of glob(3) patterns via the shell
on the remote side.
SFTP support may be enabled via a temporary scp -s flag. It is
intended for SFTP to become the default transfer mode in the
near future, at which time the -s flag will be removed. The -O
flag exists to force use of the original SCP/RCP protocol for
cases where SFTP may be unavailable or incompatible.
- sftp-server(8): add a protocol extension to support expansion of
~/ and ~user/ prefixed paths. This was added to support these
paths when used by scp(1) while in SFTP mode.
- ssh(1): add a ForkAfterAuthentication ssh_config(5) counterpart to
the ssh(1) -f flag. GHPR#231
- ssh(1): add a StdinNull directive to ssh_config(5) that allows the
config file to do the same thing as -n does on the ssh(1) command-
line. GHPR#231
- ssh(1): add a SessionType directive to ssh_config, allowing the
configuration file to offer equivalent control to the -N (no
session) and -s (subsystem) command-line flags. GHPR#231
- ssh-keygen(1): allowed signers files used by ssh-keygen(1)
signatures now support listing key validity intervals alongside
they key, and ssh-keygen(1) can optionally check during signature
verification whether a specified time falls inside this interval.
This feature is intended for use by git to support signing and
verifying objects using ssh keys.
- ssh-keygen(8): support printing of the full public key in a sshsig
signature via a -Oprint-pubkey flag.
Bugfixes
--------
* ssh(1)/sshd(8): start time-based re-keying exactly on schedule in
the client and server mainloops. Previously the re-key timeout
could expire but re-keying would not start until a packet was sent
or received, causing a spin in select() if the connection was
quiescent.
* ssh-keygen(1): avoid Y2038 problem in printing certificate
validity lifetimes. Dates past 2^31-1 seconds since epoch were
displayed incorrectly on some platforms. bz#3329
* scp(1): allow spaces to appear in usernames for local to remote
and scp -3 remote to remote copies. bz#1164
* ssh(1)/sshd(8): remove references to ChallengeResponseAuthentication
in favour of KbdInteractiveAuthentication. The former is what was in
SSHv1, the latter is what is in SSHv2 (RFC4256) and they were
treated as somewhat but not entirely equivalent. We retain the old
name as a deprecated alias so configuration files continue to work
as well as a reference in the man page for people looking for it.
bz#3303
* ssh(1)/ssh-add(1)/ssh-keygen(1): fix decoding of X.509 subject name
when extracting a key from a PKCS#11 certificate. bz#3327
* ssh(1): restore blocking status on stdio fds before close. ssh(1)
needs file descriptors in non-blocking mode to operate but it was
not restoring the original state on exit. This could cause
problems with fds shared with other programs via the shell,
bz#3280 and GHPR#246
* ssh(1)/sshd(8): switch both client and server mainloops from
select(3) to pselect(3). Avoids race conditions where a signal
may arrive immediately before select(3) and not be processed until
an event fires. bz#2158
* ssh(1): sessions started with ControlPersist were incorrectly
executing a shell when the -N (no shell) option was specified.
bz#3290
* ssh(1): check if IPQoS or TunnelDevice are already set before
overriding. Prevents values in config files from overriding values
supplied on the command line. bz#3319
* ssh(1): fix debug message when finding a private key to match a
certificate being attempted for user authentication. Previously it
would print the certificate's path, whereas it was supposed to be
showing the private key's path. GHPR#247
* sshd(8): match host certificates against host public keys, not
private keys. Allows use of certificates with private keys held in
a ssh-agent. bz#3524
* ssh(1): add a workaround for a bug in OpenSSH 7.4 sshd(8), which
allows RSA/SHA2 signatures for public key authentication but fails
to advertise this correctly via SSH2_MSG_EXT_INFO. This causes
clients of these server to incorrectly match
PubkeyAcceptedAlgorithmse and potentially refuse to offer valid
keys. bz#3213
* sftp(1)/scp(1): degrade gracefully if a sftp-server offers the
limits@openssh.com extension but fails when the client tries to
invoke it. bz#3318
* ssh(1): allow ssh_config SetEnv to override $TERM, which is
otherwise handled specially by the protocol. Useful in ~/.ssh/config
to set TERM to something generic (e.g. "xterm" instead of
"xterm-256color") for destinations that lack terminfo entries.
* sftp-server(8): the limits@openssh.com extension was incorrectly
marked as an operation that writes to the filesystem, which made it
unavailable in sftp-server read-only mode. bz#3318
* ssh(1): fix SEGV in UpdateHostkeys debug() message, triggered when
the update removed more host keys than remain present.
* many manual page fixes.
Portability
-----------
* ssh(1): move closefrom() to before first malloc. When built against
tcmalloc, the closefrom() would stomp on file descriptors created
for tcmalloc's internal use. bz#3321
* sshd(8): handle GIDs > 2^31 in getgrouplist. When compiled in 32bit
mode, the getgrouplist implementation may fail for GIDs greater than
LONG_MAX.
* ssh(1): xstrdup environment variable used by ForwardAgent. bz#3328
* sshd(8): don't sigdie() in signal handler in privsep child process;
this can end up causing sandbox violations per bz3286
The conversion from 'unsigned long' to 'int' in line 805 is due to the
laziness of declaring a carry flag as BN_ULONG, to save an extra
line of declaration.
The constants in conditional context come from the macro 'bn_cp_32'.
The unconst cast is used for initializing local BIGNUM constants; the
struct member is declared as non-const pointer.
The type widths are handled carefully, so even if there is some
conversion from 64-bit long to uint32_t, no value bits get lost.
The fallthrough case statements are a variant of Duff's Device.
The bitwise '>>' on signed value is actually on a value of type
'unsigned char', and since all platforms supported by lint have
sizeof(int) == 4, the behavior is well defined.
Fixed a problem with verifying a certificate chain when using the
X509_V_FLAG_X509_STRICT flag. This flag enables additional security
checks of the certificates present in a certificate chain. It is
not set by default.
Starting from OpenSSL version 1.1.1h a check to disallow certificates
in the chain that have explicitly encoded elliptic curve parameters
was added as an additional strict check.
An error in the implementation of this check meant that the result
of a previous check to confirm that certificates in the chain are
valid CA certificates was overwritten. This effectively bypasses
the check that non-CA certificates must not be able to issue other
certificates.
If a "purpose" has been configured then there is a subsequent
opportunity for checks that the certificate is a valid CA. All of
the named "purpose" values implemented in libcrypto perform this
check. Therefore, where a purpose is set the certificate chain will
still be rejected even when the strict flag has been used. A purpose
is set by default in libssl client and server certificate verification
routines, but it can be overridden or removed by an application.
In order to be affected, an application must explicitly set the
X509_V_FLAG_X509_STRICT verification flag and either not set a
purpose for the certificate verification or, in the case of TLS
client or server applications, override the default purpose.
([CVE-2021-3450])
Tomasz Mraz
Fixed an issue where an OpenSSL TLS server may crash if sent a
maliciously crafted renegotiation ClientHello message from a client.
If a TLSv1.2 renegotiation ClientHello omits the signature_algorithms
extension (where it was present in the initial ClientHello), but
includes a signature_algorithms_cert extension then a NULL pointer
dereference will result, leading to a crash and a denial of service
attack.
A server is only vulnerable if it has TLSv1.2 and renegotiation
enabled (which is the default configuration). OpenSSL TLS clients
are not impacted by this issue. ([CVE-2021-3449])
Peter Kaestle and Samuel Sapalski
christos